gpu-dnn-install

Author: Iacopo Poli

Description: Scripts and instructions to install CUDA, cuDNN and two of the most common deep learning frameworks (Theano and Torch).

PREREQUISITES:

1 - Download CUDA 8.0 deb(network) file for your system here. If you're using Ubuntu 16.04 on NV6, the file should be called

cuda-repo-ubuntu1604_8.0.61-1_amd64.deb

2 - Download CuDNN 5.1 for CUDA 8.0 Linux here. You have to register first and accept the License. The file should be called

cudnn-8.0-linux-x64-v5.1.tar

It should work the same with new versions of CuDNN.

All the other files needed are in this repository.

INSTALLATION

NOTE: You have to set the permission to execute the installation script files. You can do that with

chmod a+x <filename>

0 - Run this and check that it prints something, otherwise there is no NVIDIA hardware available.

lspci | grep -i nvidia

Sample output:

1 - Run

./cuda-install.sh

2 - add /usr/local/cuda-8.0/bin to PATH environment variable in .profile in home directory using nano or vim

export PATH="$PATH:/usr/local/cuda-8.0/bin"

3 - add /usr/local/cuda-8.0/lib64\ to LD_LIBRARY_PATH environment variable in .profile using nano or vim

export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda-8.0/lib64"

For points 2 & 3 you can look at the example file in this repository.

4 - Activate the changes using

source .profile

5 - Reboot the system

sudo reboot

6 - Reconnect to the machine via ssh and write a .theanorc file in the home directory equal to this. Then run the following command and check that is using gpu. It should also print a message that cuDNN is not available.

./theano-install.sh

Output:

7 - Run the following command.

./cudnn-install.sh

If you installed Theano, you can run python gpu-test.py and you should see cuDNN is now available. Output:

8 - Install Torch

./torch-install.sh

Answer yes to anything on the terminal. At the end, enter

source ~/.bashrc

9 - Install Tensorflow (GPU version)

./tensorflow-install.sh

Check that the GPU is being used by running

python tensorflow-gpu.py

10 - Install Keras by running

sudo pip install keras

When using Tensorflow backend (default setting), the code runs on GPU automatically if one is detected.

For any question you can contact me on Twitter @iacopo_poli.